圖像檢索系統(tǒng)外文翻譯--在flickr網(wǎng)站上的基于文本和基于內(nèi)容的圖像檢索 演示（英文）

1、Text-Based and Content-Based Image Retrieval on Flickr: DEMOJuan Manuel Barrios, Diego D´ ?az-Espinoza, and Benjamin BustosDepartment of Computer ScienceUniversity of ChileSantiago, ChileEmail: jbarrios,ddiaz,bebust

2、os@dcc.uchile.clAbstract—We present an image retrieval system based on acombined search of text and content. The idea is to use the text present in title, description, and tags of the images for improving the results obt

3、ained with a standard content-based search. The system contains two different user interfaces: a sidebar for the browser designed for end users, where the user must enter the Flickr URL that is visiting and the system re

4、trieves similar images from the collection, and an advanced search designed for experienced users, where the distance functions and weights can be customized.Keywords-Image databases; Information retrieval;I. INTRODUCTIO

5、NText-based search provides results with semantic sim-ilarity, while content-based search provides results with visual similarity. Due to the independence between these approaches, is likely that their combination could

6、improve the performance of a search system by benefiting of both approaches. In the present work, we present an image retrieval system based on a combined search of text and content.II. SYSTEM OVERVIEWThe present image s

7、earch system has been implementedusing Java 1.6, C++, and PostgreSQL. The set of image object were taken from Flickr web site1 using the SAPIR collection [3].In the offline phase, the images are downloaded fromFlickr usi

8、ng the URL provided by the SAPIR collection. The content-based descriptors extracted from the images were: Color Histogram 3 × 3 × 3 using RGB color space (a 27d vector), Gabor Wavelet (a 48d vector), Efficient

9、 Color Descriptor (ECD) 8 × 1 using RGB color space (a 32d vector), ECD 8 × 1 using HSV color space (a 32d vector), and Edge Local 4 × 4 (a 80d vector). The Color Histogram and Gabor Wavelet descriptors we

10、re implemented in C++ with the OpenCV library, and the other ones were implemented in Java.The text-based descriptors (title, description, and tags)were extracted from the SAPIR collection. The feature vectors were calcu

11、lated using the vectorial model and the tf-idf weighing [1]. Six feature vectors were created for1http://www.flickr.com/each image, three for the text using the Porter stemming algorithm [4] and three without stemming. A

12、 wordlist, stoplist, and the reverse file for the text features (with and without stemming) were also calculated and stored in the PostgreSQL database.In the online phase, the user enters the query image,a search text, a

13、nd a weighed distance function for each available feature. The distance functions can be metric (like Euclidean distance) or non-metric (like DPF and cosine distance). Then, the system performs a k-NN search using a weig

14、hed combination of distances, normalized by the maximum distance of a feature to the origin.All the text-based and content-based features (up to 11vectors for an image) are stored in a PostgreSQL database for efficient r

15、etrieval of a small subset, and in a binary file designed for efficient linear scan. Currently, the system contains more than 115,000 images in the collection and the binary file size is about 130 MB.The system contains

16、two different Graphical User Inter-faces: a browser sidebar (intended for end users), and an advanced search (indented for experienced users).III. BROWSER SIDEBARThe interface for end users consists in a sidebar for theb

17、rowser2 (see Figure 1). When the user is browsing images on Flickr site and wants to perform a search, he must copy the URL of the image that is viewing and paste it into the URL input text of the sidebar. The resulting

18、set of objects are displayed to the user sorted by relevance, measured as the distance of each element in the collection to the query object. The user can click on a thumbnail for accessing the published image on Flickr

19、web site.IV. ADVANCED SEARCHIn the advanced search3 (see Figure 2) the user can enterthe query image by different means: uploading an image file, entering the URL of a public image, entering a Flickr URL that displays an

20、 image (this is the only way in which the query image will contain text-based and content-based2For adding a sidebar in Firefox, create a bookmark to the URLhttp://prisma.dcc.uchile.cl/ImageSearch/sidebar.jsp, edit the p

21、roperties of the bookmark, and mark the “Load this bookmark in the sidebar” checkbox.3http://prisma.dcc.uchile.cl/ImageSearch/index.jsp2009 Second International Workshop on Similarity Search and Applications978-0-7695-37

22、65-8/09 $25.00 © 2009 IEEE DOI 10.1109/SISAP.2009.30 1552009 Second International Workshop on Similarity Search and Applications978-0-7695-3765-8/09 $25.00 © 2009 IEEE DOI 10.1109/SISAP.2009.30 1572009 Second I

23、nternational Workshop on Similarity Search and Applications978-0-7695-3765-8/09 $25.00 © 2009 IEEE DOI 10.1109/SISAP.2009.30 156Figure 1. The sidebar installed in Firefox.features), or selecting one of the previous

24、query images (useful for reusing query images). Then, the user may enter a search text. The words in the search text restricts the universe of objects for the similarity search. If the user did not enter a search text, t

25、he text-based features of the query image, if there is any, are used as search text. If the query image does not contain text-based features, the search will be performed over the whole collection.Once entered the query

26、image and text, the user canselect the features for the comparison, the distance function for each feature, and the weight of each one in the final distance. Then, the system performs a k-NN search using the normalized w

27、eighed combination of distances.The resulting set of objects are displayed to the user, andfor each result the calculated distance to the query object is shown. The system displays a “similar images” link for each result

28、, allowing to start a new search but using the selected object as a query object, and a link to the published image on Flickr web site.V. RUNTIME PERFORMANCEThe system is installed on a desktop computer with anAMD Athlon

29、 XP 1800+ and 1 GB RAM running CentOS 5.2 with Sun Java 1.6, Jakarta Tomcat 6.0 and PostgreSQL 8.1 installed. The memory usage of the VM is about 90 MB (this includes the memory used by Tomcat and our system).When the us

30、er does not enter any search text and thequery object does not contain text-based features, a linear search over all the collection is performed. In this case, the search time depends mainly on the number of compared fea

31、tures and the distance functions used. Using all the features available in the system, the search time of the linear scan is about 20 seconds. Using only the RGB Histogram 3×3×3 and Euclidean distance, the sear

32、ch time of a linear scan is about 1 second. It is worth noting that the images features are retrieved from a file during the search and are not stored in main memory.Figure 2. The advanced search.When the user enters any

33、 search text or the query containstext-based features, the inverted file is used for retrieving the set of images that contain them. In this case, on average, the search time is reduced to less than a second.VI. CONCLUSI

34、ONSIn this work, we presented a system that allows a com-bined search of text-based and content-based features on a sample of more than 115,000 images of Flickr obtained from the SAPIR collection.We are evaluating the in

35、clusion of the M3-tree [2] into thesystem as an indexing structure for dynamic combinations of metric distances.ACKNOWLEDGMENTThis work was partially supported by FONDECYT(Chile) Project 11070037.REFERENCES[1] R. A. Baez

36、a-Yates and B. Ribeiro-Neto. “Modern InformationRetrieval”. Addison-Wesley Longman Publishing Co., Inc.,Boston, MA, USA, 1999.[2] B. Bustos and T. Skopal. “Dynamic similarity search in multi-metric spaces”. In Proc. 8th

37、ACM SIGMM International Work- shop on Multimedia Information Retrieval (MIR’06), pages 137–146. ACM Press, 2006.[3] F. Falchi, M. Kacimi, Y. Mass, F. Rabitti, and P. Zezula.“SAPIR: Scalable and distributed image searchin

38、g”. In SAMT (Posters and Demos), vol. 300 of CEUR Workshop Proceed- ings, pages 11–12, Genoa, Italy, November 2007.[4] M. F. Porter. “An algorithm for suffix stripping”. Pro-gram, 14(3):130-137, July 1980. http://tartaru